To ensure that our model-based synthesis approach was effective in constraining the shared features among image pairs, we fed the final image pairs (Figure 1B) back through the neural network that generated them (GoogleNet/Inception; Szegedy et al., 2015; middle), as well as through an additional architecture (VGG19; Simonyan and Zisserman, 2014; right) to ensure that the similarity space produced was not dependent on the characteristics of one specific model. We extracted features from these networks for each of the generated images, and computed representational similarity matrices (RSMs) at the targeted layers to ensure that they reflected the intended similarity space. (A) Intended (left) and actual (center) correlations (r) of features across paired images as a function of similarity level and layer of the Inception model. We aimed for and achieved uniform correlations across pairs in lower and middle layers (2D0-4C), and linearly increasing correlations across pairs in higher layers (4D-5B, see panel B). Also shown are the actual feature correlations across layers of the alternate VGG19 model (right). In this alternate architecture (VGG19), the three highest model layers (B5P-FC2) mirrored the intended similarity for the highest model layers in the generating network (r(62) = 0.861, 0.849, 0.856 for the three highest model layers). The four low/middle layers (B1P-B4P) did as well (r(62) = 0.592, 0.542, 0.396, 0.455 for the four low/middle layers), but to a lesser extent (Steiger test ps < .001), and with lower variability across pairs (SD = 0.050, 0.017, 0.025, 0.052 for the four low/middle layers; for comparison, for the three highest layers: 0.116, 0.164, 0.169). (B) Comparison of feature correlations in each of the targeted higher layers of Inception (4D-5B). In the four highest layers (4D-5B), across all eight pairs of each of the eight endpoint axes (64 pairs total), the intended and actual feature correlations were strongly associated (r(62) = 0.970, 0.983, 0.977, 0.985, respectively, for the four highest layers). Considering each endpoint axis separately, the minimum feature correlations across the eight pairs of that axis remained high (r(6) = 0.945, 0.965, 0.966, 0.967, respectively, for the four highest layers). In the lower and middle layers (2D0-4C), feature correlations did not vary across pairs. This cannot be quantified by relating intended and actual feature correlations, given the lack of variance in intended feature correlations across pairs. Instead, the average differences between any two pairs in these eight lower/middle layers were small (M = 0.012, 0.011, 0.012, 0.041, 0.025, 0.045, 0.043, 0.053 for the eight lower/middle layers; for comparison, for the four highest layers M = 0.14, 0.19, 0.22, 0.23). The standard deviations of the feature correlations across pairs in these layers were also low (SD = 0.010, 0.010, 0.011, 0.035, 0.022, 0.039, 0.039, 0.048 for the eight lower/middle layers; for comparison, for the highest four layers SD = 0.199, 0.244, 0.259, 0.243).